AI and the World's Hard Problems · Physics · arXiv · 2026-10-26
Unmodeled states and uncertain action outcomes in agentic scanning tunneling microscopy
Tip conditioning on a scanning tunneling microscope (STM) has traditionally required skilled human experts. The researchers built quailbot, an agent harness that places an LLM inside the instrument's feedback loop. After a brief apprenticeship with an expert, frontier LLM agents completed STM tip conditioning end to end on their own and passed an independent check, though hidden or unmodeled experimental states exposed limits to their autonomy.